Vyra Dragon v1 — Semi-Realistic Dragon Character | SDXL LoRA

Welcome to my first LoRA!

Vyra Dragon is a character LoRA with a stylistic influence, trained to capture the appearance of a female dragon. It also carries some of the training dataset’s semi-realistic rendering, scale textures, and surface detail.

The original character has black, white, and gold scales, blue eyes, and horns. You can explore other colors, poses, and environments through prompting, although the original features may still influence the result. The LoRA can also provide an additional stylistic touch in other dragon-themed generations.

Quick start

• Architecture: SDXL.

• Training checkpoint: XAVIER v1.0, listed on Civitai under Illustrious.

• Trigger word: vyradrgn.

• Suggested starting strength: 0.8.

• Recommended starting range: 0.8–1.0.

• Range tested in my setups: 0.5–1.5.

Start with XAVIER v1.0, the checkpoint used for training. Results on other SDXL checkpoints can differ in character appearance, colors, anatomy, and rendering style.

Choosing the strength

In my tests, useful results were obtained from 0.5 up to 1.5. Values below 0.5 may still have an effect, but are outside the range I currently recommend.

Start at 0.8 and adjust to taste. Values around 1.0–1.5 can be useful when you want a stronger influence. At 1.5, usable results are still possible in my tested setups; above this, artifacts can appear. The best value depends on the checkpoint, prompt, seed, and other components of the workflow.

CLIP Skip and checkpoint settings

For XAVIER v1.0, the checkpoint author recommends CLIP Skip 2.

In my Civitai tests, CLIP Skip 1 did not work as expected with XAVIER, but worked with the Amanatsu checkpoint I tested. Follow the recommendations for your chosen checkpoint and adjust based on the results.

CLIP Skip and the LoRA Loader’s strength_clip are separate settings. CLIP Skip selects the text encoder output layer; strength_clip controls any text encoder weights supplied by a LoRA.

Vyra Dragon v1 was trained on the U-Net only and contains no text encoder LoRA weights. Use strength_clip: 0.0 in ComfyUI. Changing this value should not change Vyra Dragon’s effect in the standard loader.

Using Vyra Dragon in ComfyUI

1. Place the .safetensors file in ComfyUI/models/loras and refresh the model list or restart ComfyUI.

2. Load your checkpoint with Load Checkpoint. Connect its MODEL and CLIP outputs to Load LoRA, then select Vyra Dragon.

3. Set strength_model: 0.8 and strength_clip: 0.0.

4. Connect the LoRA Loader’s MODEL output to your sampler. Use its CLIP output for both positive and negative prompt encoding. If you use CLIP Set Last Layer with XAVIER, set stop_at_clip_layer to -2 and route its output to both prompt encoders.

5. Add vyradrgn to your positive prompt and generate.

To compare against the base checkpoint, set strength_model to 0.0. Keep the seed, prompt, checkpoint, and other generation settings unchanged.

Starting generation settings for XAVIER

A practical starting resolution is 1024 × 1024; portrait formats such as 832 × 1216 can also be used.

The checkpoint author recommends the following sampling settings:

• Sampler: DPM++ 2M SDE, GPU-seeded (dpmpp_2m_sde_gpu in ComfyUI).

• Scheduler: Karras.

• Steps: 30.

• CFG: 3.0.

• CLIP Skip: 2.

These are starting settings for XAVIER. The settings used in individual examples may differ. Other checkpoints should use their own recommended settings.

Simple example prompt

vyradrgn, solo, thin female dragon, black scales, white scales, gold scales, blue eyes, horns, portrait, three-quarter view, solid gray background

Optional negative prompt:

blurry, low quality, deformed, bad anatomy, extra limbs

Additional quality or style tags are optional. A simple prompt makes it easier to see the contribution of the LoRA. For full-body images, replace the portrait framing with full body, entire body visible.

Training information

• Training software: kohya_ss.

• Training checkpoint: XAVIER v1.0 (xavier_v10.safetensors).

• Dataset: 29 AI-generated images with text captions.

• Dataset coverage: portraits and full-body views from different angles.

• Training resolution: 1024 target resolution with aspect-ratio bucketing; buckets used were 832 × 1216 and 1024 × 1024.

• Training steps: 1,500.

• Network rank / dim: 32.

• Network alpha: 16.

• Learning rate: 0.0001.

• Optimizer: AdamW.

• Mixed precision: BF16.

• Training target: U-Net only; text encoder training disabled.

Examples and credits

Credit to r12x5s for XAVIER, the checkpoint used for training.

My broader generation workflow also uses the following resources:

• Lung Easter Dragon IL.

• Unfazed Soft Detailer — IL v2.0.

• ILXL Realism Slider V.1.

• Scales, Feathers and Cracks XL.

These additional LoRAs are optional. You can use Vyra Dragon with a compatible checkpoint on its own.

Examples made with additional LoRAs, ControlNet, upscaling, or other processing show the combined result of the complete workflow. For evaluating Vyra Dragon’s contribution, use a simple setup and compare it with strength set to 0. Refer to the resources and settings provided with each example.

Limitations and feedback

With a dataset of 29 images, character consistency can vary across checkpoints and compositions. Original colors or other visual features may carry over, and complex poses or anatomy may require adjustment.

Feedback and examples are welcome! If you share a result, please include the checkpoint and version, LoRA strength, CLIP Skip, and prompt when possible. That will help me understand how Vyra Dragon behaves in different setups.